The Wall Street Journal reporter Keach Hagey published a lengthy investigative report, systematically revealing for the first time the decade-long personal feud between the founders of Anthropic and OpenAI through extensive interviews with current and former employees and those close to executives at both companies. Shaping the global AI landscape is not only a battle over technological routes but also a private wound that has never healed.

Dario Amodei’s internal rhetoric in recent months has been far more intense than in public. He compared Sam Altman and Elon Musk’s legal dispute to “Hitler versus Stalin,” called OpenAI President Greg Brockman’s $25.00 million donation to a pro-Trump super political action committee “evil,” and likened OpenAI and other competitors to “tobacco companies that knowingly sell harmful products.” After the Pentagon dispute escalated, he again called OpenAI “mendacious” on Slack, writing, “These facts demonstrate a pattern of behavior I have often seen in Sam Altman.”

Internally, Anthropic calls this branding strategy creating a “healthy alternative” to competitors, and an ad during this year’s Super Bowl that indirectly satirized OpenAI’s embedding of ads in chatbots was its public manifestation.

The story begins in the living room of a shared house on Delano Street in San Francisco in 2016. Dario and his sister Daniela Amodei lived there, and OpenAI co-founder Brockman often visited because of his personal relationship with Daniela. One day, Brockman, Dario, and Daniela’s then-fiancé, effective altruism philanthropist Holden Karnofsky, sat together arguing about the correct path for AI development: Brockman believed that all Americans should be informed about what was happening on the AI frontier, while Dario and Karnofsky believed that sensitive information should be reported to the government first rather than broadcast to the public. This difference later became a watershed moment in the philosophical direction of the two companies.

Impressed by OpenAI’s talent pool, Dario joined in mid-2016, working with Brockman to train AI agents to play video games late into the night. But after four years of working together, conflicts over power and belonging intensified. In 2017, Musk, then OpenAI’s main investor, demanded a list of each employee’s contributions and laid off employees accordingly. 10% to 20% of the approximately 60-person team were dismissed one by one, which Dario considered cruel. One of those laid off later became a co-founder of Anthropic. In the same year, an ethics consultant hired by Dario proposed that OpenAI act as a coordinating entity between AI companies and the government. Brockman extrapolated from this the idea of “selling AGI to the UN Security Council’s nuclear powers,” which Dario considered almost treasonous and once considered resigning over.

After Musk left in 2018, Altman took over leadership. He and Dario reached a consensus: employees lacked confidence in the leadership of Brockman and Chief Scientist Ilya Sutskever. Dario agreed to stay on the condition that the two no longer be in charge, but soon discovered that Altman had simultaneously promised the latter two that they had the right to fire him, two contradictory promises. After the GPT series development began, the most intense conflict erupted among executives over who could participate in the language model project. Dario, then Director of Research, did not allow Brockman to get involved, and Daniela, who co-led the project with Alec Radford, threatened to resign as head to get her way. Radford’s personal wishes were drawn into the proxy war between executives.

Dario’s seniority rose with the success of GPT-2 and GPT-3, but he felt that Altman downplayed his contributions. When Brockman went on a podcast to discuss the OpenAI charter, Dario was angry because he had contributed more to the charter but was not invited; he was also unhappy to learn that Brockman and Altman were going to meet with former President Obama but had excluded him.

The conflict completely escalated during a meeting room confrontation. Altman called the Amodei siblings into the meeting room and accused them of encouraging colleagues to submit negative feedback about him to the board. The two denied it. Altman said the information came from another executive, and Daniela immediately called that executive in to confront him, but the person said he knew nothing about it. Altman then denied saying this, and the two sides argued fiercely. In early 2020, Altman asked executives to write peer reviews for each other. Brockman wrote a strongly worded feedback accusing Daniela of abusing her power and using bureaucratic processes to exclude dissidents. Altman reviewed it in advance and commented “tough but fair.” Daniela refuted it point by point, and the argument escalated to the point where Brockman once offered to withdraw the comments.

At the end of 2020, a team centered around Dario decided to leave, with Daniela leading negotiations with lawyers on departure matters. Altman personally went to Dario’s home to persuade him to stay, and Dario proposed that he would only accept reporting directly to the board and made it clear that he could not work with Brockman. Before leaving, he wrote a long memo, dividing AI companies into “market-oriented” and “public interest-oriented” types, believing that the ideal ratio was 75% public interest and 25% market. Weeks later, Dario, Daniela, and nearly a dozen employees left OpenAI to found Anthropic.

Five years later, both companies are valued at over $300.00 billion and are vying to be the first to IPO. At the closing group photo of the New Delhi AI Summit in February this year, Indian Prime Minister Modi and the technology leaders present raised their hands, while Amodei and Altman chose not to participate, only awkwardly bumping elbows.

[The Wall Street Journal]